14 struct io_completion_data {
17 int error; /* output */
18 unsigned long bytes_done[2]; /* output */
19 struct timeval time; /* output */
23 * The ->file_map[] contains a map of blocks we have or have not done io
24 * to yet. Used to make sure we cover the entire range in a fair fashion.
26 static int random_map_free(struct fio_file *f, const unsigned long long block)
28 unsigned int idx = RAND_MAP_IDX(f, block);
29 unsigned int bit = RAND_MAP_BIT(f, block);
31 dprint(FD_RANDOM, "free: b=%llu, idx=%u, bit=%u\n", block, idx, bit);
33 return (f->file_map[idx] & (1 << bit)) == 0;
37 * Mark a given offset as used in the map.
39 static void mark_random_map(struct thread_data *td, struct io_u *io_u)
41 unsigned int min_bs = td->o.rw_min_bs;
42 struct fio_file *f = io_u->file;
43 unsigned long long block;
44 unsigned int blocks, nr_blocks;
47 block = (io_u->offset - f->file_offset) / (unsigned long long) min_bs;
48 nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
50 busy_check = !(io_u->flags & IO_U_F_BUSY_OK);
53 unsigned int this_blocks, mask;
54 unsigned int idx, bit;
57 * If we have a mixed random workload, we may
58 * encounter blocks we already did IO to.
64 if ((td->o.ddir_seq_nr == 1) && !random_map_free(f, block))
67 idx = RAND_MAP_IDX(f, block);
68 bit = RAND_MAP_BIT(f, block);
70 fio_assert(td, idx < f->num_maps);
72 this_blocks = nr_blocks;
73 if (this_blocks + bit > BLOCKS_PER_MAP)
74 this_blocks = BLOCKS_PER_MAP - bit;
77 if (this_blocks == BLOCKS_PER_MAP)
80 mask = ((1U << this_blocks) - 1) << bit;
82 if (!(f->file_map[idx] & mask))
86 } while (this_blocks);
91 f->file_map[idx] |= mask;
92 nr_blocks -= this_blocks;
93 blocks += this_blocks;
97 if ((blocks * min_bs) < io_u->buflen)
98 io_u->buflen = blocks * min_bs;
101 static unsigned long long last_block(struct thread_data *td, struct fio_file *f,
104 unsigned long long max_blocks;
105 unsigned long long max_size;
107 assert(ddir_rw(ddir));
110 * Hmm, should we make sure that ->io_size <= ->real_file_size?
112 max_size = f->io_size;
113 if (max_size > f->real_file_size)
114 max_size = f->real_file_size;
116 max_blocks = max_size / (unsigned long long) td->o.ba[ddir];
124 * Return the next free block in the map.
126 static int get_next_free_block(struct thread_data *td, struct fio_file *f,
127 enum fio_ddir ddir, unsigned long long *b)
129 unsigned long long min_bs = td->o.rw_min_bs;
132 i = f->last_free_lookup;
133 *b = (i * BLOCKS_PER_MAP);
134 while ((*b) * min_bs < f->real_file_size &&
135 (*b) * min_bs < f->io_size) {
136 if (f->file_map[i] != (unsigned int) -1) {
137 *b += ffz(f->file_map[i]);
138 if (*b > last_block(td, f, ddir))
140 f->last_free_lookup = i;
144 *b += BLOCKS_PER_MAP;
148 dprint(FD_IO, "failed finding a free block\n");
152 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
153 enum fio_ddir ddir, unsigned long long *b)
155 unsigned long long r;
159 r = os_random_long(&td->random_state);
160 dprint(FD_RANDOM, "off rand %llu\n", r);
161 *b = (last_block(td, f, ddir) - 1)
162 * (r / ((unsigned long long) OS_RAND_MAX + 1.0));
165 * if we are not maintaining a random map, we are done.
167 if (!file_randommap(td, f))
171 * calculate map offset and check if it's free
173 if (random_map_free(f, *b))
176 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
181 * we get here, if we didn't suceed in looking up a block. generate
182 * a random start offset into the filemap, and find the first free
187 f->last_free_lookup = (f->num_maps - 1) *
188 (r / (OS_RAND_MAX + 1.0));
189 if (!get_next_free_block(td, f, ddir, b))
192 r = os_random_long(&td->random_state);
196 * that didn't work either, try exhaustive search from the start
198 f->last_free_lookup = 0;
199 return get_next_free_block(td, f, ddir, b);
202 static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
203 enum fio_ddir ddir, unsigned long long *b)
205 if (get_next_rand_offset(td, f, ddir, b)) {
206 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
207 f->file_name, f->last_pos, f->real_file_size);
214 static int get_next_seq_block(struct thread_data *td, struct fio_file *f,
215 enum fio_ddir ddir, unsigned long long *b)
217 assert(ddir_rw(ddir));
219 if (f->last_pos < f->real_file_size) {
220 *b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir];
227 static int get_next_block(struct thread_data *td, struct io_u *io_u,
228 enum fio_ddir ddir, int rw_seq, unsigned long long *b)
230 struct fio_file *f = io_u->file;
233 assert(ddir_rw(ddir));
237 ret = get_next_rand_block(td, f, ddir, b);
239 ret = get_next_seq_block(td, f, ddir, b);
241 io_u->flags |= IO_U_F_BUSY_OK;
243 if (td->o.rw_seq == RW_SEQ_SEQ) {
244 ret = get_next_seq_block(td, f, ddir, b);
246 ret = get_next_rand_block(td, f, ddir, b);
247 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
248 if (f->last_start != -1ULL)
249 *b = (f->last_start - f->file_offset)
250 / td->o.min_bs[ddir];
255 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
264 * For random io, generate a random new block and see if it's used. Repeat
265 * until we find a free one. For sequential io, just return the end of
266 * the last io issued.
268 static int __get_next_offset(struct thread_data *td, struct io_u *io_u)
270 struct fio_file *f = io_u->file;
271 unsigned long long b;
272 enum fio_ddir ddir = io_u->ddir;
275 assert(ddir_rw(ddir));
277 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
279 td->ddir_seq_nr = td->o.ddir_seq_nr;
282 if (get_next_block(td, io_u, ddir, rw_seq_hit, &b))
285 io_u->offset = b * td->o.ba[ddir];
286 if (io_u->offset >= f->io_size) {
287 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
288 io_u->offset, f->io_size);
292 io_u->offset += f->file_offset;
293 if (io_u->offset >= f->real_file_size) {
294 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
295 io_u->offset, f->real_file_size);
302 static int get_next_offset(struct thread_data *td, struct io_u *io_u)
304 struct prof_io_ops *ops = &td->prof_io_ops;
306 if (ops->fill_io_u_off)
307 return ops->fill_io_u_off(td, io_u);
309 return __get_next_offset(td, io_u);
312 static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u)
314 const int ddir = io_u->ddir;
315 unsigned int uninitialized_var(buflen);
316 unsigned int minbs, maxbs;
319 assert(ddir_rw(ddir));
321 minbs = td->o.min_bs[ddir];
322 maxbs = td->o.max_bs[ddir];
327 r = os_random_long(&td->bsrange_state);
328 if (!td->o.bssplit_nr[ddir]) {
329 buflen = 1 + (unsigned int) ((double) maxbs *
330 (r / (OS_RAND_MAX + 1.0)));
337 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
338 struct bssplit *bsp = &td->o.bssplit[ddir][i];
342 if (r <= ((OS_RAND_MAX / 100L) * perc))
346 if (!td->o.bs_unaligned && is_power_of_2(minbs))
347 buflen = (buflen + minbs - 1) & ~(minbs - 1);
350 if (io_u->offset + buflen > io_u->file->real_file_size) {
351 dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen,
359 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
361 struct prof_io_ops *ops = &td->prof_io_ops;
363 if (ops->fill_io_u_size)
364 return ops->fill_io_u_size(td, io_u);
366 return __get_next_buflen(td, io_u);
369 static void set_rwmix_bytes(struct thread_data *td)
374 * we do time or byte based switch. this is needed because
375 * buffered writes may issue a lot quicker than they complete,
376 * whereas reads do not.
378 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
379 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
382 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
387 r = os_random_long(&td->rwmix_state);
388 v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
389 if (v <= td->o.rwmix[DDIR_READ])
395 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
397 enum fio_ddir odir = ddir ^ 1;
401 assert(ddir_rw(ddir));
403 if (td->rate_pending_usleep[ddir] <= 0)
407 * We have too much pending sleep in this direction. See if we
412 * Other direction does not have too much pending, switch
414 if (td->rate_pending_usleep[odir] < 100000)
418 * Both directions have pending sleep. Sleep the minimum time
419 * and deduct from both.
421 if (td->rate_pending_usleep[ddir] <=
422 td->rate_pending_usleep[odir]) {
423 usec = td->rate_pending_usleep[ddir];
425 usec = td->rate_pending_usleep[odir];
429 usec = td->rate_pending_usleep[ddir];
431 fio_gettime(&t, NULL);
432 usec_sleep(td, usec);
433 usec = utime_since_now(&t);
435 td->rate_pending_usleep[ddir] -= usec;
438 if (td_rw(td) && __should_check_rate(td, odir))
439 td->rate_pending_usleep[odir] -= usec;
445 * Return the data direction for the next io_u. If the job is a
446 * mixed read/write workload, check the rwmix cycle and switch if
449 static enum fio_ddir get_rw_ddir(struct thread_data *td)
454 * see if it's time to fsync
456 if (td->o.fsync_blocks &&
457 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
458 td->io_issues[DDIR_WRITE] && should_fsync(td))
462 * see if it's time to fdatasync
464 if (td->o.fdatasync_blocks &&
465 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
466 td->io_issues[DDIR_WRITE] && should_fsync(td))
467 return DDIR_DATASYNC;
470 * see if it's time to sync_file_range
472 if (td->sync_file_range_nr &&
473 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
474 td->io_issues[DDIR_WRITE] && should_fsync(td))
475 return DDIR_SYNC_FILE_RANGE;
479 * Check if it's time to seed a new data direction.
481 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
483 * Put a top limit on how many bytes we do for
484 * one data direction, to avoid overflowing the
487 ddir = get_rand_ddir(td);
489 if (ddir != td->rwmix_ddir)
492 td->rwmix_ddir = ddir;
494 ddir = td->rwmix_ddir;
495 } else if (td_read(td))
500 td->rwmix_ddir = rate_ddir(td, ddir);
501 return td->rwmix_ddir;
504 void put_file_log(struct thread_data *td, struct fio_file *f)
506 int ret = put_file(td, f);
509 td_verror(td, ret, "file close");
512 void put_io_u(struct thread_data *td, struct io_u *io_u)
516 io_u->flags |= IO_U_F_FREE;
517 io_u->flags &= ~IO_U_F_FREE_DEF;
520 put_file_log(td, io_u->file);
523 if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
525 flist_del_init(&io_u->list);
526 flist_add(&io_u->list, &td->io_u_freelist);
528 td_io_u_free_notify(td);
531 void clear_io_u(struct thread_data *td, struct io_u *io_u)
533 io_u->flags &= ~IO_U_F_FLIGHT;
537 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
539 struct io_u *__io_u = *io_u;
541 dprint(FD_IO, "requeue %p\n", __io_u);
545 __io_u->flags |= IO_U_F_FREE;
546 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(__io_u->ddir))
547 td->io_issues[__io_u->ddir]--;
549 __io_u->flags &= ~IO_U_F_FLIGHT;
550 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
552 flist_del(&__io_u->list);
553 flist_add_tail(&__io_u->list, &td->io_u_requeues);
558 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
560 if (td->io_ops->flags & FIO_NOIO)
563 io_u->ddir = get_rw_ddir(td);
566 * fsync() or fdatasync() or trim etc, we are done
568 if (!ddir_rw(io_u->ddir))
572 * See if it's time to switch to a new zone
574 if (td->zone_bytes >= td->o.zone_size) {
576 io_u->file->last_pos += td->o.zone_skip;
577 td->io_skip_bytes += td->o.zone_skip;
581 * No log, let the seq/rand engine retrieve the next buflen and
584 if (get_next_offset(td, io_u)) {
585 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
589 io_u->buflen = get_next_buflen(td, io_u);
591 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
595 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
596 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
597 dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset,
598 io_u->buflen, io_u->file->real_file_size);
603 * mark entry before potentially trimming io_u
605 if (td_random(td) && file_randommap(td, io_u->file))
606 mark_random_map(td, io_u);
609 * If using a write iolog, store this entry.
612 dprint_io_u(io_u, "fill_io_u");
613 td->zone_bytes += io_u->buflen;
618 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
647 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
649 __io_u_mark_map(td->ts.io_u_submit, nr);
650 td->ts.total_submit++;
653 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
655 __io_u_mark_map(td->ts.io_u_complete, nr);
656 td->ts.total_complete++;
659 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
663 switch (td->cur_depth) {
685 td->ts.io_u_map[index] += nr;
688 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
725 assert(index < FIO_IO_U_LAT_U_NR);
726 td->ts.io_u_lat_u[index]++;
729 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
770 assert(index < FIO_IO_U_LAT_M_NR);
771 td->ts.io_u_lat_m[index]++;
774 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
777 io_u_mark_lat_usec(td, usec);
779 io_u_mark_lat_msec(td, usec / 1000);
783 * Get next file to service by choosing one at random
785 static struct fio_file *get_next_file_rand(struct thread_data *td,
786 enum fio_file_flags goodf,
787 enum fio_file_flags badf)
793 long r = os_random_long(&td->next_file_state);
796 fno = (unsigned int) ((double) td->o.nr_files
797 * (r / (OS_RAND_MAX + 1.0)));
799 if (fio_file_done(f))
802 if (!fio_file_open(f)) {
805 err = td_io_open_file(td, f);
811 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
812 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
816 td_io_close_file(td, f);
821 * Get next file to service by doing round robin between all available ones
823 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
826 unsigned int old_next_file = td->next_file;
832 f = td->files[td->next_file];
835 if (td->next_file >= td->o.nr_files)
838 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
839 if (fio_file_done(f)) {
844 if (!fio_file_open(f)) {
847 err = td_io_open_file(td, f);
849 dprint(FD_FILE, "error %d on open of %s\n",
857 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
859 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
863 td_io_close_file(td, f);
866 } while (td->next_file != old_next_file);
868 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
872 static struct fio_file *__get_next_file(struct thread_data *td)
876 assert(td->o.nr_files <= td->files_index);
878 if (td->nr_done_files >= td->o.nr_files) {
879 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
880 " nr_files=%d\n", td->nr_open_files,
886 f = td->file_service_file;
887 if (f && fio_file_open(f) && !fio_file_closing(f)) {
888 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
890 if (td->file_service_left--)
894 if (td->o.file_service_type == FIO_FSERVICE_RR ||
895 td->o.file_service_type == FIO_FSERVICE_SEQ)
896 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
898 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
900 td->file_service_file = f;
901 td->file_service_left = td->file_service_nr - 1;
903 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
907 static struct fio_file *get_next_file(struct thread_data *td)
909 struct prof_io_ops *ops = &td->prof_io_ops;
911 if (ops->get_next_file)
912 return ops->get_next_file(td);
914 return __get_next_file(td);
917 static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
922 f = get_next_file(td);
929 if (!fill_io_u(td, io_u))
933 td_io_close_file(td, f);
935 fio_file_set_done(f);
937 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
938 td->nr_done_files, td->o.nr_files);
945 struct io_u *__get_io_u(struct thread_data *td)
947 struct io_u *io_u = NULL;
952 if (!flist_empty(&td->io_u_requeues))
953 io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
954 else if (!queue_full(td)) {
955 io_u = flist_entry(td->io_u_freelist.next, struct io_u, list);
964 assert(io_u->flags & IO_U_F_FREE);
965 io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF);
966 io_u->flags &= ~IO_U_F_TRIMMED;
969 flist_del(&io_u->list);
970 flist_add(&io_u->list, &td->io_u_busylist);
972 io_u->flags |= IO_U_F_IN_CUR_DEPTH;
973 } else if (td->o.verify_async) {
975 * We ran out, wait for async verify threads to finish and
978 pthread_cond_wait(&td->free_cond, &td->io_u_lock);
986 static int check_get_trim(struct thread_data *td, struct io_u *io_u)
988 if (td->o.trim_backlog && td->trim_entries) {
991 if (td->trim_batch) {
994 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
995 td->last_ddir != DDIR_READ) {
996 td->trim_batch = td->o.trim_batch;
998 td->trim_batch = td->o.trim_backlog;
1002 if (get_trim && !get_next_trim(td, io_u))
1009 static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1011 if (td->o.verify_backlog && td->io_hist_len) {
1014 if (td->verify_batch) {
1017 } else if (!(td->io_hist_len % td->o.verify_backlog) &&
1018 td->last_ddir != DDIR_READ) {
1019 td->verify_batch = td->o.verify_batch;
1020 if (!td->verify_batch)
1021 td->verify_batch = td->o.verify_backlog;
1025 if (get_verify && !get_next_verify(td, io_u))
1033 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1034 * etc. The returned io_u is fully ready to be prepped and submitted.
1036 struct io_u *get_io_u(struct thread_data *td)
1041 io_u = __get_io_u(td);
1043 dprint(FD_IO, "__get_io_u failed\n");
1047 if (check_get_verify(td, io_u))
1049 if (check_get_trim(td, io_u))
1053 * from a requeue, io_u already setup
1059 * If using an iolog, grab next piece if any available.
1061 if (td->o.read_iolog_file) {
1062 if (read_iolog_get(td, io_u))
1064 } else if (set_io_u_file(td, io_u)) {
1065 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1070 assert(fio_file_open(f));
1072 if (ddir_rw(io_u->ddir)) {
1073 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1074 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1078 f->last_start = io_u->offset;
1079 f->last_pos = io_u->offset + io_u->buflen;
1081 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_WRITE)
1082 populate_verify_io_u(td, io_u);
1083 else if (td->o.refill_buffers && io_u->ddir == DDIR_WRITE)
1084 io_u_fill_buffer(td, io_u, io_u->xfer_buflen);
1085 else if (io_u->ddir == DDIR_READ) {
1087 * Reset the buf_filled parameters so next time if the
1088 * buffer is used for writes it is refilled.
1090 io_u->buf_filled_len = 0;
1095 * Set io data pointers.
1097 io_u->xfer_buf = io_u->buf;
1098 io_u->xfer_buflen = io_u->buflen;
1102 if (!td_io_prep(td, io_u)) {
1103 if (!td->o.disable_slat)
1104 fio_gettime(&io_u->start_time, NULL);
1108 dprint(FD_IO, "get_io_u failed\n");
1113 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1115 const char *msg[] = { "read", "write", "sync", "datasync",
1116 "sync_file_range", "wait", "trim" };
1120 log_err("fio: io_u error");
1123 log_err(" on file %s", io_u->file->file_name);
1125 log_err(": %s\n", strerror(io_u->error));
1127 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
1128 io_u->offset, io_u->xfer_buflen);
1131 td_verror(td, io_u->error, "io_u error");
1134 static void io_completed(struct thread_data *td, struct io_u *io_u,
1135 struct io_completion_data *icd)
1138 * Older gcc's are too dumb to realize that usec is always used
1139 * initialized, silence that warning.
1141 unsigned long uninitialized_var(usec);
1144 dprint_io_u(io_u, "io complete");
1147 assert(io_u->flags & IO_U_F_FLIGHT);
1148 io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1151 if (ddir_sync(io_u->ddir)) {
1152 td->last_was_sync = 1;
1155 f->first_write = -1ULL;
1156 f->last_write = -1ULL;
1161 td->last_was_sync = 0;
1162 td->last_ddir = io_u->ddir;
1164 if (!io_u->error && ddir_rw(io_u->ddir)) {
1165 unsigned int bytes = io_u->buflen - io_u->resid;
1166 const enum fio_ddir idx = io_u->ddir;
1167 const enum fio_ddir odx = io_u->ddir ^ 1;
1170 td->io_blocks[idx]++;
1171 td->io_bytes[idx] += bytes;
1172 td->this_io_bytes[idx] += bytes;
1174 if (idx == DDIR_WRITE) {
1177 if (f->first_write == -1ULL ||
1178 io_u->offset < f->first_write)
1179 f->first_write = io_u->offset;
1180 if (f->last_write == -1ULL ||
1181 ((io_u->offset + bytes) > f->last_write))
1182 f->last_write = io_u->offset + bytes;
1186 if (ramp_time_over(td)) {
1187 unsigned long uninitialized_var(lusec);
1189 if (!td->o.disable_clat || !td->o.disable_bw)
1190 lusec = utime_since(&io_u->issue_time,
1192 if (!td->o.disable_lat) {
1193 unsigned long tusec;
1195 tusec = utime_since(&io_u->start_time,
1197 add_lat_sample(td, idx, tusec, bytes);
1199 if (!td->o.disable_clat) {
1200 add_clat_sample(td, idx, lusec, bytes);
1201 io_u_mark_latency(td, lusec);
1203 if (!td->o.disable_bw)
1204 add_bw_sample(td, idx, bytes, &icd->time);
1205 if (__should_check_rate(td, idx)) {
1206 td->rate_pending_usleep[idx] =
1207 ((td->this_io_bytes[idx] *
1208 td->rate_nsec_cycle[idx]) / 1000 -
1209 utime_since_now(&td->start));
1211 if (__should_check_rate(td, idx ^ 1))
1212 td->rate_pending_usleep[odx] =
1213 ((td->this_io_bytes[odx] *
1214 td->rate_nsec_cycle[odx]) / 1000 -
1215 utime_since_now(&td->start));
1218 if (td_write(td) && idx == DDIR_WRITE &&
1220 td->o.verify != VERIFY_NONE)
1221 log_io_piece(td, io_u);
1223 icd->bytes_done[idx] += bytes;
1226 ret = io_u->end_io(td, io_u);
1227 if (ret && !icd->error)
1230 } else if (io_u->error) {
1231 icd->error = io_u->error;
1232 io_u_log_error(td, io_u);
1234 if (td->o.continue_on_error && icd->error &&
1235 td_non_fatal_error(icd->error)) {
1237 * If there is a non_fatal error, then add to the error count
1238 * and clear all the errors.
1240 update_error_count(td, icd->error);
1247 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1250 if (!td->o.disable_clat || !td->o.disable_bw)
1251 fio_gettime(&icd->time, NULL);
1256 icd->bytes_done[0] = icd->bytes_done[1] = 0;
1259 static void ios_completed(struct thread_data *td,
1260 struct io_completion_data *icd)
1265 for (i = 0; i < icd->nr; i++) {
1266 io_u = td->io_ops->event(td, i);
1268 io_completed(td, io_u, icd);
1270 if (!(io_u->flags & IO_U_F_FREE_DEF))
1276 * Complete a single io_u for the sync engines.
1278 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1279 unsigned long *bytes)
1281 struct io_completion_data icd;
1283 init_icd(td, &icd, 1);
1284 io_completed(td, io_u, &icd);
1286 if (!(io_u->flags & IO_U_F_FREE_DEF))
1290 td_verror(td, icd.error, "io_u_sync_complete");
1295 bytes[0] += icd.bytes_done[0];
1296 bytes[1] += icd.bytes_done[1];
1303 * Called to complete min_events number of io for the async engines.
1305 int io_u_queued_complete(struct thread_data *td, int min_evts,
1306 unsigned long *bytes)
1308 struct io_completion_data icd;
1309 struct timespec *tvp = NULL;
1311 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1313 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1318 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1320 td_verror(td, -ret, "td_io_getevents");
1325 init_icd(td, &icd, ret);
1326 ios_completed(td, &icd);
1328 td_verror(td, icd.error, "io_u_queued_complete");
1333 bytes[0] += icd.bytes_done[0];
1334 bytes[1] += icd.bytes_done[1];
1341 * Call when io_u is really queued, to update the submission latency.
1343 void io_u_queued(struct thread_data *td, struct io_u *io_u)
1345 if (!td->o.disable_slat) {
1346 unsigned long slat_time;
1348 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1349 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen);
1354 * "randomly" fill the buffer contents
1356 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1357 unsigned int max_bs)
1359 if (!td->o.zero_buffers)
1360 fill_random_buf(io_u->buf, max_bs);
1362 memset(io_u->buf, 0, max_bs);